Claude SDK POC

Multi-agent research orchestrator using the Claude Agent SDK. Spawns parallel specialized agents to gather codebase context for implementation planning.

Usage

# New interactive session
bun run dev                          # Interactive prompt
bun run dev "your research query"    # With initial query
bun run dev --cwd /path/to/repo      # Custom working directory

# Continue/Resume sessions
bun run dev --continue               # Continue most recent session
bun run dev --resume SESSION_ID      # Resume specific session by ID
bun run dev:continue                 # Shortcut for --continue

# Headless/piped mode (runs query and exits)
echo "your query" | bun run dev      # Pipe input
cat query.txt | bun run dev          # From file

Interactive Session

After the initial query, the session stays open for follow-ups:

<session_id> > your follow-up question here

Exiting

• Type EXIT and press Enter
• Press Ctrl+C
• Press Ctrl+D

Session Continuation

• Session IDs are displayed when a session starts
• Use --resume SESSION_ID to resume any previous session
• Use --continue to resume the most recent session
• Both continuation modes fork the session (original is preserved)

Headless Mode

When stdin is not a TTY (piped input), the tool runs in headless mode:

• Runs the query and prints the result to stdout
• Prints session ID to stderr (for scripting/resume)
• Exits after the first response (no interactive prompt)

Structure

• index.ts - CLI entry point (Commander.js)
• src/interactive.ts - Interactive session handler
• src/research.ts - Research orchestrator (legacy one-shot mode)
• src/agents/ - 4 specialized research agents

Key Files

| File | Purpose | |------|---------| | src/interactive.ts | Interactive session with AsyncIterable prompt | | src/research.ts | Legacy planning1Research() function | | src/agents/index.ts | Agent exports | | src/evals/run-evals.ts | Eval runner with WorkflowFn type |

If Working On...

Interactive mode: See src/interactive.ts. Uses SDK's AsyncIterable prompt mode for multi-turn conversations.

Evals: See src/evals/. Run bun run eval, accept with bun run eval:accept, diff with bun run eval:diff.

New workflow function: Export from src/research.ts matching (input: ResearchInput) => Promise<ResearchResult>, then add eval cases in run-evals.ts.

Agents: See src/agents/. Each agent is an AgentDefinition with description, tools, model, prompt.

References

• SDK Documentation - Claude TypeScript SDK usage
• Implementation Plan - Full planning doc